An AI agent that is unsure as to how to proceed in any given situation has two options. It can either say so and wait for the user to decide (usually by asking either a single question or presenting a series of questions and/options), or it can decide itself based on its “best guess”. The second option is far more likely because that way it can quickly produce a confident, well-formed answer. Nothing in the output shows that it made this “best guess”. When asked, however, it will readily admit that it was simply making up an answer based on what it “thought ” was correct information.
Here’s an example of what happens when it does that. It will take a prompt like “sort these documents and put the details in a spreadsheet”, which reads as a complete instruction even when it is actually incomplete. The instruction says nothing about which details to sort, in what order, in what format, and what to do with a receipt too faint to read. These are four crucial details not covered in the initial instruction, and now the agent tries to work out what it should do all by itself. Instead of coming back and asking for more clarity, it simply proceeds based on what it “thinks” is the correct way to carry out the instruction based upon what it knows from its training data. In other words, it thinks that it knows how your instruction would normally be carried out, and proceeds based upon that information. This may or may not be at all what you actually wanted.
Nothing about that is a failure of intelligence. The instruction was incomplete, the work could not proceed without the gaps being filled, and it filled the gaps itself.
How to give clear instructions
When someone sees this happening their first instinct is to write more firmly. Capital letters, warnings, a longer paragraph of insistence. In fact, it is not unknown to see people using voice input speaking with a raised voice or even shouting at the AI agent who they simply can’t believe is not understanding what they’re trying to tell it! This does not work, and the reason is worth remembering: firmness doesn’t actually clarify the instruction. An instruction repeated more loudly contains exactly the same ambiguities as before.
What does close the gaps is settling the question in advance. Every question the job would otherwise force the agent to answer on your behalf is one that should be answered in advance, once, in writing. The more the whole task can be specified the better. Explain what is important and what finished looks like, what to leave alone, and what to do with the cases it cannot handle.
Such a detailed instruction that specifies these things is called a “brief” in this course, and it belongs to the first of the five parts of the framework. However a brief is not just a better prompt. It’s a set of instructions that, when given at the beginning of a session, will be used by AI as a session proceeds to guide its decision making.
Clear instructions are important, but an Agent also needs to know that to work on
A proper brief is only one part of what the agent needs to do a good job. The brief is part of the agent’s context: it tells the agent what to do. The other part tells it what to work on, which is the facts and information it needs access to in order to carry out the task correctly. So that’s what the next section of this course will be focusing on – creating a proper brief and giving the AI agent the knowledge that it needs to work correctly.
The example this course builds
“Learning by doing” is for many people a really good way to take on board the skills needed to work with AI. It’s not enough to just understand the theory, you also need to work with practical examples to see how AI actually works. We’ve prepared a set of sample data that you can use during the course to learn how to practically use AI. You’ll learn how to direct AI to carry out a number of sophisticated tasks and learn how to build the framework to ensure that the AI agent returns more and more reliable results.
These documents are an example that you can use to learn methods that fit any task that repeats and has to come out right. It doesn’t matter what your data is, the methods and techniques are the same.
So welcome to the course. We hope you enjoy it and find it useful in your journey of exploration and learning with this new technology. Using these methods and techniques will ensure that you can get significantly more useful, reliable and safe results from Agentic AI.